Fully automated quantitative cephalometry using convolutional neural networks

Abstract

Quantitative cephalometry plays an essential role in clinical diagnosis, treatment, and surgery. Development of fully automated techniques for these procedures is important to enable consistently accurate computerized analyses. We study the application of deep convolutional neural networks (CNNs) for fully automated quantitative cephalometry for the first time. The proposed framework utilizes CNNs for detection of landmarks that describe the anatomy of the depicted patient and yield quantitative estimation of pathologies in the jaws and skull base regions. We use a publicly available cephalometric x-ray image dataset to train CNNs for recognition of landmark appearance patterns. CNNs are trained to output probabilistic estimations of different landmark locations, which are combined using a shape-based model. We evaluate the overall framework on the test set and compare with other proposed techniques. We use the estimated landmark locations to assess anatomically relevant measurements and classify them into different anatomical types. Overall, our results demonstrate high anatomical landmark detection accuracy ([Formula: see text] to 2% higher success detection rate for a 2-mm range compared with the top benchmarks in the literature) and high anatomical type classification accuracy ([Formula: see text] average classification accuracy for test set). We demonstrate that CNNs, which merely input raw image patches, are promising for accurate quantitative cephalometry.

Keywords: artificial neural networks; feed-forward neural networks; image recognition; machine vision; predictive models; statistical learning; supervised learning; x-ray applications.

Fig. 1

Schematics of the proposed cephalometric landmark detection framework.

Fig. 2

Domains of true and false landmark samples for training.

Fig. 3

CNN architecture for local information based probability mapping.

Fig. 4

Four test images and corresponding estimations of the probability of candidate location being the landmark. The estimations are shown for three example landmarks: L1, L10, and L19. Red crosses in test images denote the ground truth locations of L1 (top), L10 (bottom), and L19 (middle). Inside each red dashed box, the probability of each point being the corresponding landmark is shown by the grayscale color map such that the white corresponds to 0 and black corresponds to 1. The red dashed boxes are centered at ground-truth landmark locations and have a size of $4\times 4{\mathrm{cm}}^2$.

Fig. 5

Success detection rates of the technique proposed in this article versus other benchmarks for test sets from IEEE ISBI 2014 Challenge and IEEE ISBI 2015 Challenge Datasets.